Configure hyperparameters from the CLI (Advanced)

Audience: Users looking to modularize their code for a professional project.

Pre-reqs: You must have read (Mix models and datasets).

As a project becomes more complex, the number of configurable options becomes very large, making it inconvenient to control through individual command line arguments. To address this, CLIs implemented using LightningCLI always support receiving input from configuration files. The default format used for config files is YAML.

Tip

If you are unfamiliar with YAML, it is recommended that you first read What is a yaml config file?.


Run using a config file

To run the CLI using a yaml config, do:

python main.py fit --config config.yaml

Individual arguments can be given to override options in the config file:

python main.py fit --config config.yaml --trainer.max_epochs 100

Automatic save of config

To ease experiment reporting and reproducibility, by default LightningCLI automatically saves the full YAML configuration in the log directory. After multiple fit runs with different hyperparameters, each one will have in its respective log directory a config.yaml file. These files can be used to trivially reproduce an experiment, e.g.:

python main.py fit --config lightning_logs/version_7/config.yaml

The automatic saving of the config is done by the special callback SaveConfigCallback. This callback is automatically added to the Trainer. To disable the save of the config, instantiate LightningCLI with save_config_callback=None.

Tip

To change the file name of the saved configs to e.g. name.yaml, do:

cli = LightningCLI(..., save_config_kwargs={"config_filename": "name.yaml"})

It is also possible to extend the SaveConfigCallback class, for instance to additionally save the config in a logger. An example of this is:

class LoggerSaveConfigCallback(SaveConfigCallback):
    def save_config(self, trainer: Trainer, pl_module: LightningModule, stage: str) -> None:
        if isinstance(trainer.logger, Logger):
            config = self.parser.dump(self.config, skip_none=False)  # Required for proper reproducibility
            trainer.logger.log_hyperparams({"config": config})


cli = LightningCLI(..., save_config_callback=LoggerSaveConfigCallback)

Tip

If you want to disable the standard behavior of saving the config to the log_dir, then you can either implement __init__ and call super().__init__(*args, save_to_log_dir=False, **kwargs) or instantiate the LightningCLI as:

cli = LightningCLI(..., save_config_kwargs={"save_to_log_dir": False})

Note

The save_config method is only called on rank zero. This allows to implement a custom save config without having to worry about ranks or race conditions. Since it only runs on rank zero, any collective call will make the process hang waiting for a broadcast. If you need to make collective calls, implement the setup method instead.


Prepare a config file for the CLI

The --help option of the CLIs can be used to learn which configuration options are available and how to use them. However, writing a config from scratch can be time-consuming and error-prone. To alleviate this, the CLIs have the --print_config argument, which prints to stdout the configuration without running the command.

For a CLI implemented as LightningCLI(DemoModel, BoringDataModule), executing:

python main.py fit --print_config

generates a config with all default values like the following:

seed_everything: null
trainer:
  logger: true
  ...
model:
  out_dim: 10
  learning_rate: 0.02
data:
  data_dir: ./
ckpt_path: null

Other command line arguments can be given and considered in the printed configuration. A use case for this is CLIs that accept multiple models. By default, no model is selected, meaning the printed config will not include model settings. To get a config with the default values of a particular model would be:

python main.py fit --model DemoModel --print_config

which generates a config like:

seed_everything: null
trainer:
  ...
model:
  class_path: lightning.pytorch.demos.boring_classes.DemoModel
  init_args:
    out_dim: 10
    learning_rate: 0.02
ckpt_path: null

Tip

A standard procedure to run experiments can be:

# Print a configuration to have as reference
python main.py fit --print_config > config.yaml
# Modify the config to your liking - you can remove all default arguments
nano config.yaml
# Fit your model using the edited configuration
python main.py fit --config config.yaml

Configuration items can be either simple Python objects such as int and str, or complex objects comprised of a class_path and init_args arguments. The class_path refers to the complete import path of the item class, while init_args are the arguments to be passed to the class constructor. For example, your model is defined as:

# model.py
class MyModel(L.LightningModule):
    def __init__(self, criterion: torch.nn.Module):
        self.criterion = criterion

Then the config would be:

model:
  class_path: model.MyModel
  init_args:
    criterion:
      class_path: torch.nn.CrossEntropyLoss
      init_args:
        reduction: mean
    ...

LightningCLI uses jsonargparse under the hood for parsing configuration files and automatic creation of objects, so you don’t need to do it yourself.

Note

Lightning automatically registers all subclasses of LightningModule, so the complete import path is not required for them and can be replaced by the class name.

Note

Parsers make a best effort to determine the correct names and types that the parser should accept. However, there can be cases not yet supported or cases for which it would be impossible to support. To somewhat overcome these limitations, there is a special key dict_kwargs that can be used to provide arguments that will not be validated during parsing, but will be used for class instantiation.

For example, then using the lightning.pytorch.profilers.PyTorchProfiler profiler, the profile_memory argument has a type that is determined dynamically. As a result, it’s not possible to know the expected type during parsing. To account for this, your config file should be set up like this:

trainer:
  profiler:
    class_path: lightning.pytorch.profilers.PyTorchProfiler
    dict_kwargs:
      profile_memory: true

Compose config files

Multiple config files can be provided, and they will be parsed sequentially. Let’s say we have two configs with common settings:

# config_1.yaml
trainer:
  num_epochs: 10
  ...

# config_2.yaml
trainer:
  num_epochs: 20
  ...

The value from the last config will be used, num_epochs = 20 in this case:

$ python main.py fit --config config_1.yaml --config config_2.yaml

Use groups of options

Groups of options can also be given as independent config files. For configs like:

# trainer.yaml
num_epochs: 10

# model.yaml
out_dim: 7

# data.yaml
data_dir: ./data

a fit command can be run as:

$ python main.py fit --trainer trainer.yaml --model model.yaml --data data.yaml [...]